본문 바로가기
Weekly · AI Ecosystem Briefing

Nov 1 – Nov 7, 2026

Korean original

The AI contest is no longer decided by model performance but by verified work delivered per dollar. Chinese open-weight models and Google's TPU camp control that cost curve, so America's closed frontier labs can defend margins only through specialisation and regulatory friendliness.

US companies send about 90% of their open-source model spending to Chinese firms, and Kimi alone takes 61%. Purchasing decisions are being made on price, not geopolitics. In the same period Gemini 4 Carbon delivered Claude Opus 5.5-class performance at lower cost, and a Moonshot open-weight model beat OpenAI on a financial-research benchmark. The premium for general-purpose performance is thinning fast. Pressure is building on the demand side too. Ramp data show enterprise AI spending has stalled, and on SemiAnalysis’s real-work benchmark even the best agent fails to clear 12%. Inflated expectations for agents are giving way to a phase of ROI scrutiny. The closed labs are responding by leaving the general-purpose race for specialist agents such as GPT-6 Cyber. Its generation of patches for 948 kernel privilege-escalation vulnerabilities suggests security will be the first area to make money. In infrastructure, NVIDIA held HBM capacity growth to 3% while promising up to 30 times more agent throughput per dollar, and Broadcom TPU compute doubled to roughly 2 million H100 equivalents. The axis of competition has moved from FLOPS to memory and total cost of ownership. Meanwhile, voter support for data-centre limits, the Democrats’ signalled sweeping AI bill and OpenAI’s power-grid programme all point to permits and power, not chips, as the bottleneck on expansion. China is closing its own stack from chips to distribution, running DeepSeek V5 on Huawei silicon and putting Qwen into Apple Intelligence. The next two weeks will be decided by pricing responses and regulatory moves.

Key moves

Predictions (with triggers)

At least one of OpenAI, Google and Anthropic officially announces an API price cut or a new low-cost tier for its flagship model · Next 2 weeks
Notices of changes to input or output token prices on each company's official pricing page or blog
DeepSeek V5 does not ship as a formal open-weight release at 8 trillion parameters and stays at the reported preparation stage · Next 2 weeks
Whether V5 weights or a technical report appear on DeepSeek's official channels or Hugging Face
The top public score on SemiAnalysis's real-work benchmark stays in the 12% range, with no agent result above 15% · Next 2 weeks
Whether a score of 15% or higher is listed on the SemiAnalysis leaderboard or in follow-up reporting

Watchlist

Based on 249 items over 7 days

SubscribePast issues